Start From Scratch: Turning an Old Tower Into Your First Homelab

Everyone’s first homelab article online seems to assume you already own a rack, three nodes, and a managed switch. Mine started with none of that. It started with a desktop tower that was already sitting under the desk doing nothing important and a stubborn refusal to spend money before I’d proven I’d actually use the thing.

That refusal turned out to be the most useful rule I set. This post is the “GBP0 start” - what a normal old PC can genuinely do, in what order I’d do it again, and the honest moment where reused hardware stops being enough and you have to decide whether you’re serious.

The only hardware question that matters at the start

You do not need to buy anything to begin. You need to answer one question: does the machine you already own have enough RAM to hold a couple of small services in memory at once?

That is the real bottleneck for a first homelab, not the CPU. An older quad-core chip is fine. What runs out first is memory. My rough rule:

• 8 GB - enough to learn on. One or two light containers. You will feel it the moment you try anything AI-related.
• 16 GB - the comfortable starting line. Several containers, or one small local language model, without thrashing.
• 32 GB+ - now you can run a local model and other services and stop thinking about it.

If your tower is sitting at 8 GB, a RAM upgrade is the single highest-value pound you can spend, and often the only thing you need to buy for months. Match the speed and generation your motherboard already takes rather than guessing.

Where you’ll hit a wall: if the tower still boots from an old mechanical hard drive, everything will feel slow in a way that makes you blame the wrong thing. A small SSD for the OS and containers is usually the second-best upgrade after RAM.

Step 1: Do not reinstall anything yet

My tower was running Windows 11. The internet’s instinct is to wipe it and install Linux immediately. Do not do that on day one. Windows 11 ships with WSL2 - a real Linux kernel running alongside Windows - and it is the gentlest possible on-ramp. You get a genuine Ubuntu environment without committing to wiping anything.

wsl --install

Reboot, pick a username and password, and you have Ubuntu.

This matters for confidence. The fastest way to quit a new hobby is to brick your only working machine in week one. WSL2 lets you back out of everything by closing a window.

Step 2: Docker, because everything useful comes as a container

Almost every self-hosted service worth running ships as a Docker image. Install Docker Desktop on Windows, enable WSL2 integration, and you can run that whole world from the Ubuntu terminal.

docker version
docker ps

The first time docker ps returns an empty table instead of an error, you have cleared the only setup hurdle that stops most people. Everything after this is just pulling images.

One genuinely annoying gotcha: Docker Desktop on WSL2 stores its data wherever it defaults to, and moving that data later is not a casual drag-and-drop. Decide early which drive has the space and let Docker put its data there from the start.

Step 3: The payoff is a local AI model on hardware you already owned

This is the moment a homelab stops being abstract. You can run a real language model, locally, offline, for free, on that same old tower.

The stack I started with, and still the one I would recommend to a beginner, is three pieces:

Piece	Job
Ollama	Runs the model locally.
Open WebUI	Gives you a chat interface in the browser.
A small model	llama3.1:8b for general chat, gemma:2b when you want faster responses.

ollama pull llama3.1:8b
ollama pull gemma:2b

Point Open WebUI at Ollama, open it in a browser, and you are chatting with a model that lives entirely on your hardware. On a CPU-only old tower it will not be fast, but it works, and “it works on the thing I already owned” is the entire point of starting here.

Where you’ll hit a wall: an 8B model is about the ceiling for comfortable CPU-only inference. If you want bigger models or faster responses, that becomes a GPU conversation.

What the GBP0 start actually taught me

Three things shaped everything after:

1. Reuse first, buy second. I proved I would use a homelab before I spent real money.
2. The bottleneck is rarely the CPU. It is RAM, then disk speed, then much later the GPU.
3. Every “I’ll move that later” decision is a trap. Docker data location, storage layout, and drive roles are cheap to decide early and painful to undo later.

Where this goes next

A single tower running a few containers is a fantastic start and a terrible final destination. The wall I hit was not compute. It was the network. The moment I had services I actually relied on, plus smart-home devices and everyone else’s hardware sharing one flat network, things got messy quickly.

That is the next part of this series: moving from “one PC doing everything” to a properly segmented home network with entry-level managed gear.

---

RigForge is reader-supported. Where I link hardware in later parts, those may be affiliate links. The default advice here is still the honest one: buy nothing until the reused hardware shows you the next real limitation.